Vast.ai: GPU auction marketplace, RTX 4090 from USD 0.31/h, A100 from USD 0.40/h, globally distributed

What is Vast.ai?

Vast.ai is an auction marketplace for GPU compute, founded in 2018 and registered in San Francisco. The business model differs fundamentally from RunPod, Lambda Labs or CoreWeave: Vast.ai operates no own hardware but brokers GPU hours between private and professional providers (sellers) and renters (buyers). As of May 2026 there are over 10,000 available GPU instances from globally distributed data centres, home setups and rack hostings – geographically spread across North America, Europe and Asia, with a strong weight in the US and Eastern Europe.

The auction model works on bid logic. Sellers set a minimum price per GPU hour, buyers choose available cards by filter (GPU type, VRAM, RAM, storage, region, bandwidth, verification). There are two modes: on-demand instances with guaranteed runtime (seller cannot interrupt), and interruptible instances at markedly lower price (seller can outbid at any time and terminate the instance).

May 2026 inventory: NVIDIA RTX 3090 (24 GB) from USD 0.18/h interruptible, RTX 4090 (24 GB) from USD 0.31/h interruptible up to USD 0.45/h on-demand, RTX A6000 (48 GB) from USD 0.55/h, A100-40GB from USD 0.40/h interruptible, A100-80GB from USD 0.78-1.20/h, H100-80GB from USD 1.85-2.50/h, H200 from USD 3.50/h. Prices fluctuate daily with availability – one of the lowest H100 prices on the market, with all the consequences of an auction platform.

Provider verification is a key category. "Verified" providers are audited by Vast.ai (hardware, network, SSH access). "Datacenter" providers operate from real data centres with standard infrastructure. "Community" providers are less strictly audited, often home hosting on consumer internet. Anyone processing production or sensitive data filters on "Verified" and "Datacenter".

The platform is API-first with CLI (`vastai`), Python SDK and a web console. Docker container-based workloads are standard – anyone with an image can start it. Pre-built templates for Stable Diffusion, vLLM, Ollama, Jupyter are available.

Why it matters

In May 2026 Vast.ai is the cheapest GPU cloud on the market – factor 2-3 below RunPod Community Cloud, factor 3-5 below Lambda Labs, factor 5-10 below hyperscalers. Three use cases justify the use: experimental workloads, hyperparameter search and non-sensitive training jobs.

Experimental workloads: an ML researcher tests 20 different model architectures over 2 hours each. On Vast.ai with RTX 4090 at USD 0.31/h that costs USD 12.40 for 40 GPU hours. At RunPod that would be USD 13.60, at AWS around USD 80, at Scaleway EUR 24. The saving is marginal per job, but cumulatively significant over a research quarter.

Hyperparameter search: a job that tests 100 different configurations via Bayesian optimisation benefits massively from parallel auction availability. 100x parallel A100 hours on Vast.ai are in many cases available and cheaper than at any other provider. The interruptible logic is no problem here – failed hyperparameter trials are restarted, the search algorithm is resilient.

Training jobs without personal reference: open-weight model finetuning on publicly available data (Wikipedia, Common Crawl, code repositories) has no personal-data compliance requirement. Filtering on "Verified Datacenter" yields usable stability at the lowest prices. With checkpoint logic (save every 30 minutes), interruptible pods are robustly usable.

The downside is clear. Vast.ai sellers are heterogeneous – from serious data centres to hobby hosts with a single RTX 3090 in the living room. That means: no guaranteed data residency, no ISO certification of the platform, no DPA in standardised form, no compliance package for regulated mandates. For Swiss fiduciaries with client data, Vast.ai is not suitable.

Regulatory position: Vast.ai is a US Delaware LLC, subject to US law and the CLOUD Act. Data sits, depending on the seller, in arbitrary countries – the region filters help with geographic restriction, but the legal position of the individual seller may vary. A transfer impact assessment for Vast.ai is complicated and in May 2026 not defensible for most Swiss compliance audiences.

How it works

Ordering: via the web console vast.ai or the CLI `vastai`. Account creation by email verification, payment by credit card or crypto (Bitcoin, USDC), prepaid balance in USD. Instance search via filters: GPU type, VRAM, RAM, storage, region, seller verification, bandwidth, disk IO. Sorting typically by USD/h ascending.

Sample prices May 2026 (auction-dependent, fluctuating): RTX 3090 24GB interruptible USD 0.18-0.25/h, on-demand USD 0.32-0.45/h. RTX 4090 24GB interruptible USD 0.31-0.40/h, on-demand USD 0.45-0.65/h. A100-40GB interruptible USD 0.40-0.60/h, on-demand USD 0.80-1.20/h. A100-80GB interruptible USD 0.78-1.10/h, on-demand USD 1.20-1.80/h. H100-80GB interruptible USD 1.85-2.50/h, on-demand USD 2.50-3.50/h. Platform fee (Vast.ai) typically 10-15% on the rent price, already included in the displayed prices.

Network: each instance has a public IPv4 address, SSH port typically 22 plus additional ports per configuration. Bandwidth filters allow selection by download/upload speed – anyone streaming large datasets should filter on "Datacenter" providers with guaranteed bandwidth.

Storage: each instance has local SSD/NVMe storage. Cloud sync to S3-compatible buckets (own or provided by the seller) is the typical workflow for data persistence. Vast.ai offers no own persistent volume service beyond the pod lifecycle.

Contract details: no classic contract, only platform terms of service. DPA is not included in the standard setup – the seller is a different party than Vast.ai, the renter is the end customer, which is a 3-party constellation. A genuine data processing agreement per nFADP Art. 9 with a standardised contract is therefore complex and in May 2026 not mappable via Vast.ai.

Docker and templates: each workload runs as a Docker container. Pre-built templates for ComfyUI, Stable Diffusion, A1111, vLLM, Ollama, Jupyter, ML Studio are available in the platform. Own containers can be pulled via Docker Hub, GitHub Container Registry or other registries.

Interruption logic: with interruptible instances another bidder can outbid the instance – the running workload typically gets 30-120 seconds warning time (SIGTERM) before SIGKILL. Anyone not implementing checkpointing loses at most the last few minutes of work. Anyone implementing it correctly is robust against interruptions.

Migration: typically no migration needed – Vast.ai is job-based, not a permanent hoster.

Vast.ai setup for hyperparameter search in 5 steps

1. 01Create an account at vast.ai, top up the prepaid balance, install the CLI `vastai` locally and configure the API key.
2. 02Set filters: "Verified Datacenter", desired GPU (e.g. A100-40GB), region (EU or North America), minimum bandwidth and storage size.
3. 03Prepare Docker image: own workload or template (vLLM, ComfyUI, Jupyter), upload to Docker Hub or GitHub Container Registry.
4. 04Start job: `vastai search offers` to filter, `vastai create instance` to start the best card, configure an S3 bucket for checkpointing.
5. 05Start the workload with checkpoint logic (save model state every 30 minutes), automatic restart script on the next instance on interruption, terminate the instance and export data at the end.

When to use Vast.ai

Vast.ai is the right choice for (a) extremely cost-sensitive experiments, (b) hyperparameter search with many parallel trials, or (c) training workloads on publicly available data without personal reference. Concrete cases: ML research group with a USD 50,000 quarterly budget. Open-source AI project with community-funded finetuning. Startup with experimental workloads testing several architectures per day.

For Stable Diffusion and ComfyUI workloads Vast.ai is also popular. RTX 3090 or RTX 4090 at USD 0.18-0.31/h is the cheapest option on the market for ComfyUI generations. With checkpoint logic and cloud backup to S3, the interruption-risk solution is pragmatic.

For hyperparameter sweeps with Ray Tune or Optuna Vast.ai is very fitting. Distributing 50-200 GPU hours in parallel over a few days, with stop logic on poor trials, yields costs far below standard cloud providers.

For LLM benchmark studies (latency tests, throughput measurements across hardware) Vast.ai is the simplest platform for GPU diversification. A range of H100, A100, L40S and A6000 instances can be started within half an hour.

When not to use

Vast.ai is not appropriate for (a) data with personal reference under nFADP, (b) professional secrecy per Art. 321 SCC, (c) bank or insurance data under FINMA regulation, or (d) productive inference workloads with availability requirements.

The multi-party constellation (renter – Vast.ai – seller) makes a clean data processing agreement difficult. A TIA for Vast.ai with unclear provider location is in May 2026 not realistic to run. For Swiss fiduciaries with client data, Exoscale Zurich, Infomaniak or another Swiss AG is the clean choice.

Productive inference with guaranteed availability: because sellers can terminate their instances without warning (interruptible) or the hardware is not 24/7 mission-critical capable, Vast.ai is not suitable for productive live inference with availability requirements. Even on-demand instances have no genuine hardware SLA.

Anyone with a training job with a hard deadline and guaranteed end time (e.g. for a conference paper with submission date) should switch to RunPod Secure Cloud or Lambda Labs. Vast.ai availability can theoretically be reduced at critical moments.

Anyone expecting professional support with callback guarantee and response times is in the wrong place at Vast.ai. Support is community-oriented (Discord, forum) and not at enterprise SLA level.

General caveat: when choosing an unverified Community provider, the theoretical risk exists that the provider has SSH access to the instance. Anyone processing even moderately sensitive data should filter on "Verified Datacenter".

Trade-offs

FAQ

Related topics

Sources

1. Vast.ai – Pricing and instance marketplace · 2026-05
2. Vast.ai – Verified Datacenter program · 2026-04
3. Vast.ai – CLI documentation and API · 2026-05
4. Vast.ai – Interruptible vs on-demand modes · 2026-05